Automatic interpretation of digital maps

ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing journal homepage: www.elsevier.com/locate/isprsjprs

Automatic interpretation of digital maps Volker Walter ∗ , Fen Luo Institute for Photogrammetry, University of Stuttgart, Geschwister-Scholl-Str. 24D, D-70174 Stuttgart, Germany

article

info

Article history: Received 27 April 2010 Received in revised form 22 February 2011 Accepted 23 February 2011 Available online 21 March 2011 Keywords: Interpretation Classification Spatial data mining Analysis Recognition

abstract In the past, the availability and/or the acquisition of spatial data were often the main problems of the realization of spatial applications. Meanwhile this situation has changed: on one hand, comprehensive spatial datasets already exist and on the other hand, new sensor technologies have the ability to capture fast and with high quality large amounts of spatial data. More and more responsible for the increasing accessibility of spatial data are also collaborative mapping techniques which enable users to create maps by themselves and to make them available in the internet. However, the potential of this diversity of spatial data can only hardly be utilized. Especially maps in the internet are represented very often only with graphical elements and no explicit information about the map’s scale, extension and content is available. Nevertheless, humans are able to extract this information and to interpret maps. For example, it is possible for a human to distinguish between rural and industrial areas only by looking at the objects’ geometries. Furthermore, a human can easily identify and group map objects that belong together. Also the type, scale and extension of a map can be identified under certain conditions only by looking at the objects’ geometries. All these examples can be subsumed under the term ‘‘map interpretation’’. In this paper it is discussed how map interpretation can be automated and how automatic map interpretation can be used in order to support other processes. The different kinds of automatic map interpretation are discussed and two approaches are shown in detail. © 2011 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

1. Introduction Map interpretation can be described as the geographical interpretation of the elements of a map and the interrelationships between these elements. Maps contain much implicit information which is not stored explicitly, but which can be derived with map interpretation, such as settlement structures, city centers or population density. In the past, map interpretation was mainly a discipline for cartographers who used paper maps as a communication medium for solving spatial problems. Today, computer-based interpretation can be used to derive automatically new information that is not stored explicitly in the dataset. The applications of automatic map interpretation are manifold. Beside the task of solving spatial problems, automatic map interpretation can be used to support other applications, like map generalization (Zhang, 2004; Filippovska et al., 2008), matching of spatial datasets (Volz and Walter, 2006; Butenuth et al., 2007) data fusion (Chen and Walter, 2009; Wiemann and Bernard, 2010) or data update (Anders and Fritsch, 1996; Walter, 2004).

∗

Corresponding author. Tel.: +49 711 68584091; fax: +49 711 68583297. E-mail address: [email protected] (V. Walter).

Furthermore, automatic map interpretation can support data retrieval processes. In conjunction with search engines and digital globes, new applications are possible. For example, the company Google explores currently techniques for the automatic indexing of audio files with speech recognition software (Google, 2008). With such indexing techniques it is possible to retrieve audio files automatically in the same way as normal web pages. The same idea can be transferred to digital maps. With automatic map indexing techniques it would be possible to assign keywords to maps or spatial parts of maps. The keywords can be used to support spatial searches. For example, if a user wants to find all maps that contain a golf course with an ocean view, the corresponding maps can be found, even if this information is not stored explicitly. Automatic map interpretation techniques are also important for integrating heterogeneous data. Up to now, spatial data are captured and stored primarily in centralized structures. However, future spatial applications are characterized by the fact that they must integrate distributed heterogeneous data from different sources: ‘‘While the last 10 years of Web-based mapping have been very important, the next five years will be revolutionary as we move from simple mapping and geospatial visualization to full geoservices on the Web‘‘ (Dangermond, 2008). The integration of heterogeneous data from different sources is only possible if the

0924-2716/$ – see front matter © 2011 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved. doi:10.1016/j.isprsjprs.2011.02.010

520

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

semantic of the data is known. Map interpretation techniques can be used to enrich spatial data with metadata and to create ontologies in order to build a Semantic Web (Feigenbaum et al., 2007). Ontologies are a formal representation of knowledge with specific concepts. In Kuhn (2002) it is discussed how the semantics of categories in Geographic Information Systems can be modeled by ontologies. An overview of ontology-driven approaches for geographic data integration can be found in Buccella et al. (2009). A combination of the technologies of the Semantic Web and collaborative approaches (Web 2.0 Technologies) (Fischer, 2008) is currently being discussed as the concept of ‘‘Web 3.0’’. Even though, Web 3.0 is at the moment more an idea, realizations of collaborative mapping approaches, like for example OpenStreetMap (2010), are already available. The enrichment of spatial data of collaborative mapping portals can be done automatically with map interpretation techniques or at least semiautomatically with an automatic tagging support system. In this case, the system analyses the spatial data automatically and makes suggestions to the user, how he can tag the data with semantic information. The remaining paper is structured as follows. After a literature review, automatic map interpretation approaches are subdivided into different classes and possible solutions are outlined. Then, two different automatic map interpretation algorithms are presented: an object-based classification based on self-organizing maps and a region-based approach based on a raster-based algorithm. Finally, the results are discussed and an outlook to future research is given. 2. Related works The automatic derivation of unknown information from databases is also known under the term Data Mining or Knowledge Discovery (Frawley et al., 1991). In the context of spatial data, these techniques are also called Geographic Knowledge Discovery (GKD) or Geographic Data Mining (Miller and Han, 2009). Data mining techniques are used to derive unknown information from huge data sets that is not visible for a human. This applies only partly to this work, because we want also to derive information that is visible for humans but which is not modeled and stored explicitly in the database. On the other hand the information can be visible, but people do not see it, because the information might be cluttered by other information, e.g. a roundabout is present in a road vector set but it might not be seen immediately, when it is not highlighted. Therefore, map interpretation can be seen as a mixture between data mining and image interpretation. Automatic map interpretation has already been discussed in other works. One criterion to differentiate between different existing map interpretation approaches is the type of input data (raster or vector data). An approach for the automatic interpretation of scanned topographic maps with query languages can be found in Graeff and Carosio (2002). The interpretation here is done with pattern recognition algorithms in the raster domain. The detected objects are implicitly contained in the raster images but were explicitly modeled when the corresponding analog map was produced. Therefore, the objects are already visible, but cannot be queried because of the raster representation. A combination of a raster- and vector-based approach for automatic map interpretation is discussed in Viglino and PierrotDeseilligny (2003). The input for this process is a raster map that is converted into a vector representation. Different object classes (for example buildings, hangars or parcels) are reconstructed with low level primitive extraction and subsequent classification. Vector-based approaches are often based on techniques from the field of Artificial Intelligence (AI). For example, Sester

(2000) presents an approach for the semi-automatic interpretation of unstructured vector data based on machine learning techniques. A graph-based approach for clustering unstructured point data can be found in Anders (2003). The approach is completely parameter-free and can be applied for very different data sources. Data mining approaches in context of spatially aware search engines are discussed in Heinzle and Sester (2004). They describe the automatic extraction of classical metadata from spatial data sets and concepts of information retrieval to derive implicit information with data mining algorithms. In Heinzle et al. (2007) this work is continued and algorithms for the automatic recognition of patterns in road network data are developed. The search for patterns in maps in order to detect implicit information for the automatic map generalization is described in Mackaness and Edwards (2002). They argue that any map can be seen as a subset of possible patterns and a map generalization is a set of transformations from one pattern to another. An ontology driven pattern recognition approach for the detection of terraced houses in vector data is presented in Lüscher et al. (2008). They use ontologies to describe the characteristics of terraced houses and map this ontology onto a pattern recognition process. Steinhauer et al. (2001) present a method for the automatic interpretation of abstract regions in a map. An abstract region consists of several map objects, which are grouped to a single object. The process is subdivided into two steps. First, region candidates are selected based on an evaluation of neighborhood relations. Then, objects which consist of a hierarchical combination of single objects are recognized with a grammar-based compiler approach. The interpretation of spatial data cannot only be done with 2D data but also with 3D data. For example, Schleinkofer (2007) uses Neural Networks in order to classify building constructions. The construction elements are classified into the categories walls, doors, windows, ceilings, ceiling openings, pillars and beams by an evaluation of their spatial extension, surface area, intersection area with other objects and object coordinates. Automatic sketch interpretation is a problem which has many similarities to the problem of automatic map interpretation (Wuersch and Egenhofer, 2008). However, in sketch interpretation the main focus is more on segmentation, classification and labeling (Sezgin and Davis, 2005), whereas in map interpretation the focus is more on the following tasks, like clustering or data mining. Also the abstraction level of the input data in sketch interpretation is typically higher as in map interpretation. Maps consist typically of well-formed geometrical objects whereas sketches could also be represented by very simple geometrical entities. Nevertheless, both research areas have a large overlap. 3. Classification of map interpretation approaches Automatic map interpretation approaches can be subdivided into approaches that derive metadata (information about the map: map type, map extension, number of map elements, etc.) and approaches that interpret the map content (interpretation of objects, grouping of objects, region interpretation etc.). Another classification schema is to distinguish between approaches that are based only on an evaluation of the geometry and approaches that also use thematic data (attributes, explicit stored object classes, etc.) as an input for the interpretation. In the following, automatic interpretation techniques are subdivided into seven different classes, depending on the kind of information that should be interpreted. The different classes can depend on each other. For example, it is easier to interpret the map scale when the map type is known and vice versa. Also the map type has an influence to the representation of the map objects and vice versa.

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

521

3.1. Map recognition

3.3. Interpretation of map scale and map extension

A very basic map interpretation approach is to classify images into maps and other images. The images could be represented with raster or vector data. A simple computer-based approach is to describe the images with a feature vector and to classify this vector with an unsupervised or supervised classification. A similar method is described in Walter (2004). In this approach a GIS change detection based on an object-wise classification of remote sensing data is introduced. The approach classifies not single pixels but groups of pixels. The pixels of an aerial image are grouped by objects and different measures that are derived from the aerial image represent the feature space. Each object is described by a 16-dimensional feature vector and classified to the most likely class based on a supervised maximum likelihood classification. Some typical features, which could be used to classify maps, are for example that they have high contrast (in order to make them easy to read) and that they contain large homogenous areas. Grouping images into categories is also an important problem in content-based image retrieval. Vailaya et al. (2001) present an approach for image classification that uses a Bayesian classifier for a hierarchical classification of vacation images. At the highest level, images are classified into indoor and outdoor images. At the next level, outdoor images are classified into city and landscape images. At the lowest level, landscape images are classified into sunset, forest and mountain images. An alternative simple interpretation approach is to evaluate the name of the image, because many maps have a filename containing the word map. Furthermore, metadata of the images (if available) could be evaluated or the text in the website that contains the image. For example, an image search (on 21 July 2010) with the search engine Google founds 1,060,000,000 images for the keyword map. Not all of these images are cartographic maps because Google makes no map interpretation but searches only for the keyword map in the website. These images could be for example artificial maps that contain objects that do not exist in the real world (Fig. 1(a)) or images that contain a map that cannot be used for spatial analyses because of geometric distortions (Fig. 1(b)) or maps that contain non-cartographic information (Fig. 1(c)) or images that contain no maps at all (Fig. 1(d)). We evaluated the first 500 hits of the search for the keyword map. The result was that 386 images represented cartographic maps and 114 images contained other information. When we project this ratio to the number of all images that were found, it can be expected that the internet contains huge amounts of cartographic information.

If a map is represented with vector data in a known coordinate system or as a georeferenced raster image, the map scale and map extension are stored explicitly in the data. Otherwise, specific techniques are needed in order to derive this information. The scale can be interpreted by an evaluation of the map objects. For example, streets are represented with polygons in a large scale and with lines on a small scale. But this requires a prior map interpretation or at least a prior interpretation of single map objects (see next paragraph). If no information about the map content is available, the interpretation of the map scale is difficult. One solution for this problem is to derive the map scale from the map extension. In order to determine the map extension, map objects can be selected which have a shape that is very ‘‘eye-catching’’. Then it could be attempted to match these objects with objects of other maps, where the map extension is already known. This problem is similar to the problem of matching of spatial data from different sources. In both problems, corresponding objects in two datasets have to be found. Different approaches have been suggested to solve this problem: a method based on buffers and measures from the information theory was described in Walter and Fritsch (1999). This approach was extended by Zhang and Meng (2006) with unsymmetrical buffers. An approach for matching of linear objects in raster data was developed by Seo and O’Hara (2009). If the map scale is known, the interpretation of the map extension becomes easier because this reduces significantly the search space of the matching approach.

3.2. Interpretation of map type

3.5. Interpretation of complex map objects

Maps have a typical appearance depending on their map type. Fig. 2 shows some examples of different map types. For a human it is possible to distinguish for example between height maps, street maps or thematic maps. Even the age of a map can be at least vaguely determined. The automatic interpretation of the map type depends on the content of the map. For example roads, rivers and contour lines can be distinguished by an evaluation of the node types in the map (Heinzle and Sester, 2004). A cadastral map (at least in Germany) consists mainly of lines, a land use map of homogenous areas and a weather map is represented typically with raster cells with a high variability. The main task therefore is to find the typical characteristics of different map types in order to classify them. The following classification can be done similar to the classification of images into maps and non-maps.

Several map objects together can form a complex object that must not be stored explicitly in the database, but can be interpreted by a human. For example, a complex intersection consists of several street objects or a settlement consists of several building blocks and a building block consists of several houses (Fig. 3). The detection of complex structures can be done by an evaluation of topological and metrical neighborhoods with clustering algorithms. Clustering algorithms are used to group objects together in such a way that the objects in a group are as similar as possible and that different object groups are as dissimilar as possible with regard to a similarity measure. Clustering is a subdiscipline of data mining and cannot only be used to find complex objects that are visible for a human but also to derive interrelationships in a dataset that are hidden.

3.4. Interpretation of single map objects Different map objects have typical geometrical appearances depending on their object type. For example: houses have typically rectangular structures, rivers are normally represented with smooth lines and streets are often represented with straight lines. Some objects have a very typical unique appearance, like football stadiums or churches. In order to interpret the object type, the objects can be represented with a feature vector, which consists of different geometrical measures, and then classified with an unsupervised or supervised classification algorithm. An example of this kind of approach is shown in the next chapter. The interpretation of single map objects becomes easier, when the map scale is known, because the object appearance is depended on the map scale. Beside the object geometry, also object attributes can be evaluated. For example, an attribute value with the letters AVE, RD or ST is a hint that the object type is a street and the letters are standing for Avenue, Road or Street.

522

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

a

b

d

c

Fig. 1. Examples of results of a Google image search for the keyword map that do not represent cartographic maps.

Fig. 2. Different map types.

Fig. 3. Example of complex map objects: (left) intersection, (right) building block.

In map interpretation especially the detection of visible clusters is of importance. However, also the derivation of hidden interrelations can be used for map interpretation. For example, the interrelationship that large cities are located very often at a river is a hidden interrelationship that can be extracted automatically with data mining techniques (Heinzle and Sester, 2004). This interrelationship can also be used to find automatically large cities in a dataset. Complex objects can also be interpreted with model based approaches. Weindorf (2002) propose an approach that is based on a grammatical description of objects. The model is represented with grammatical rules in PROLOG. The inputs are geometrical

primitives (lines and text elements) which are grouped together by interpreting the grammatical rules. A graph based approach for detecting geometrical structures in road networks is described in Heinzle et al. (2005). The patterns (e.g. grids, stars and rings) are used for the automatic determination of city centers in vector maps. 3.6. Interpretation of regions Whereas the extension of a complex map object can be calculated precisely by the extension of the related single map objects, the definition of regions is broader. Regions are areas in

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

a map where the boarder is fuzzy, like industrial areas, residential areas, town center or touristic attractive areas. The interpretation of regions can be realized — similar to the interpretation of complex map objects — with clustering techniques. Also image interpretation techniques can be used. In the next chapter, an image analysis approach, which considers the fuzzy appearance of regions, is shown with an example. The concept of regions has similarities to the concept of vague places. A vague place is a specific place with fuzzy borders (e.g. South of France or Rocky Mountains). An approach to model and find the boundaries of vague places based on an evaluation of web pages can be found in Jones et al. (2008). The assumption of this approach is that web pages, that contain the name of a vague place, contain also very often places with well known coordinates. The extension of the vague place is then calculated based on a density surface modeling of the frequency of the occurrence of the colocated places. 3.7. Further map interpretation approaches Map interpretation can be defined as the deriving of new information from a map that is not explicitly stored in the data. Therefore, every classical GIS analysis can also be seen as a map interpretation. For example, a high water simulation or network analysis lead to new information that is not stored explicitly in the data. From the authors’ point of view, the main difference is that in classical GIS analyses the process chain is typically completely defined by user knowledge whereas in typical map interpretation probabilistic and machine learning techniques are applied. In practice it is difficult to define a clear distinction between these two kinds of approaches, because also mixed approaches, where user knowledge and learning techniques or probabilistic techniques are applied, can be used (see for example Malerba et al., 2003). 4. Map interpretation examples In the following, two examples of automatic map interpretation are shown. In the first example Kohonen Feature Maps are used for a classification of single map objects and in the second example a raster-based approach is realized in order to detect regions in a map. 4.1. Example 1: classification of map objects with Kohonen Feature Maps Kohonen Feature Maps are a special type of Artificial Neural Networks. Artificial Neural Networks are an approach to simulate biological neural networks. Different types of Artificial Neural Networks have been developed, like single-layer feedforward networks, multilayer-networks, recurrent networks or SelfOrganizing Maps. The Kohonen Feature Map was developed by Kohonen (1982) and is a type of a Self-Organizing Map which uses unsupervised learning in order to organize the connections between the neurons. That means that the neural network does not know what the correct classification for a specific input is. The function of the network is therefore to categorize different inputs into clusters which have similar characteristics. Besides Kohonen Feature Maps there exist other unsupervised learning techniques like k-means, competitive learning or vector quantization. In principle every unsupervised learning technique can be used for map interpretation. We use Kohonen Feature Maps because they are easy to implement, they can be trained without much effort and they are robust against noisy input data. The following realization of a Kohonen Feature Map is based on a self-developed Java program.

523

Fig. 4. Topology of a Kohonen Feature Map.

A Kohonen Feature Map consists of an input layer and an output layer (see Fig. 4). The layers are represented with neurons. Objects, that should be classified, are represented with a feature vector which describes the object characteristics. This feature vector is the input for the Kohonen Feature Map. The output is a classification of the objects into different object classes. The definition of the characteristics of the feature vector is the most important part and has to be done very carefully. The classification consists of a learning phase and a mapping phase. Each neuron of the output layer corresponds to an object type and is facilitated or inhibited by the other neurons. Each object type has a center of activation on the output layer which represents the most appropriate neuron of this object type. Around the center of activation are other neurons which correspond to the same object type, but with a decreasing activation the greater the distance to this center is. The search for the center of activation for each object type is the main part of the learning phase. A detailed description of the mathematical basics of Kohonen Feature Maps can be found in Agarwal and Skupin (2008). In the following, the practical use of Kohonen Feature Maps for the classification of map objects is discussed. We tested the classification approach on two test areas (one rural and one urban area) in the scale of 1:10,000. The objects that should be classified in the rural area are: buildings, stadiums, roundabouts, highways, major roads, side roads and rail tracks (see Fig. 6). First, the object characteristics of the different object classes have to be defined with a vector consisting of 0 and 1 values. Altogether ten object characteristics are used (see Table 1) which results in an input vector with the dimension 10. The selection of appropriate object characteristics is the most important part of the definition of a Kohonen Feature Map. Typically, a good configuration can only be found by testing and optimizing different combinations of possible characteristics. For different types of input data (for example maps in different scales) different optimal characteristics must be defined. If the input data is similar (same type of data, but other location) the configuration must not be changed. In the next step, the numbers of neurons of the Kohonen Feature Map have to be defined. The number of neurons of the input layer corresponds to the number of different object characteristics. The optimum number of neurons of the output layer was estimated by testing. Based on an evaluation of different configurations, we use 30*30 neurons in the output layer. After the training phase, the neurons of the output layer of the Kohonen Feature Map can be colored depending on the object type which they are representing (see Fig. 5). The classification result of the urban area is shown in Fig. 6. The recognition rate is 79%. The incorrect classified objects are marked with red circles. It can be seen that in some cases small streets were classified as buildings or large buildings were

524

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

Table 1 Definition of the object characteristics for urban areas.

Object area < 800 m2 Number of right angles of object ≥ 3 800 m2 < object area < 15 000 m2 Object perimeter > 1000 m Ratio object parameter to object width > 10 m 35 m < object width < 40 m 25 m < object width < 35 m Object width < 18 m Object completely embedded in another object Number of intersections with other objects ≥ 2

Building

Stadium

Round-about

Highway

Major road

Side road

Railtrack

1 1 1 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 1 0

0 0 1 0 0 0 0 0 1 1

0 0 0 1 1 1 0 0 0 0

0 0 0 1 1 0 1 0 0 0

0 0 1 0 1 0 0 1 0 1

0 0 1 1 1 0 0 1 0 1

Table 2 Definition of the object characteristics for rural areas.

Number of right angles of object ≥ 3 Object area < 800 m2 Object area > 300 000 m2 Object perimeter > 2000 m Ratio object perimeter to object width > 10 25 m < objectwidth < 35 m Object width < 18 m Number of intersections with other objects ≥ 2

Building

Side road

Major road

Agricultural area

Forest

1 1 0 0 0 0 0 0

0 0 0 0 1 0 1 1

0 0 0 1 1 1 0 0

0 0 0 0 0 0 0 0

0 0 1 1 0 0 0 0

Fig. 5. Output layer of the Kohonen Feature Map for urban areas after the training phase.

classified as streets. Also two rail tracks were classified wrongly as side roads. In a second test we used a rural area as input for the classification. Since rural areas contain other object classes than urban areas we defined a different configuration. The object classes that should be classified in the rural area are: buildings, side roads, major roads, agricultural areas and forest. The object characteristics for rural areas are defined in Table 2. The best classification result could be achieved with a Kohonen Feature Map with 20*20 neurons in the output layer. Fig. 7 shows the output layer after the training phase. The classification result is shown in Fig. 8. The recognition rate in the rural area is 94%. The wrong classified objects are marked with red circles. Three small agricultural areas were classified wrongly as side roads and one large building was classified wrongly as an agricultural area. The reason for the lower recognition rate in urban areas is that the objects’ characteristics in urban areas have a higher variability as in rural areas and that more different object classes were used. However, these are only first results and it can be expected that the recognition rate can be further improved. 4.2. Example 2: region-based classification with a raster-based algorithm At the beginning of the region-based classification, an operator can define two different parameters for generating the regions:

Fig. 6. Classification example of an urban area. Incorrect classified objects are marked with red circles.

the grid cell size of the resulting raster map and the radius around the center of each grid cell (influence radius), so that the area for which the region indicators have to be observed can be calculated (area of influence). After the operator has chosen these parameters, the area is subdivided into equally sized, square-

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

Fig. 7. Output layer of the Kohonen Feature Map for rural areas after the training phase.

The calculation of the node density indicator can easily be done by counting all topological nodes which are within the cluster’s radius. The calculation of the rectangularity indicator is somewhat more difficult to obtain. For the calculation we only use nodes with more than two incident edges that are within the cluster radius. Nodes with only two incident edges do not represent intersections. Therefore we do not use them for the calculation of the rectangularity. Typically, nodes like that are used to model an attribute change of a road element between two intersections (for example change of the maximum speed). For nodes with more than two incident edges we calculate the n − 1 smallest angles between these incident edges. The values of these angles are then normalized onto an interval between 0 and 90: normalized_angle =

Fig. 8. Classification example of a rural area. Incorrect classified objects are marked with red circles.

shaped grid cells. Then, using the center point of each cell, the area of influence is determined and different indicators are calculated for each grid cell. The result is a raster layer for each indicator. As indicators for recognizing different levels of urbanity we use node density and rectangularity of streets, since we assume that (at least in Germany) in city centers are more topological nodes and more irregular, non-orthogonal streets than in suburbs or rural areas.

525

angle 180 − angle

if angle between (0,90) if angle between (90,180).

Angles that are larger than 180 degrees must not be considered because we evaluate only the n − 1 smallest angles between the n incident edges of a node. Finally, the rectangularity of each node is calculated by the arithmetic mean of all normalized angles. The closer this value is to 90, the higher is the degree of rectangularity of the corresponding node. The rectangularity indicator is finally calculated by the arithmetic mean of the rectangularity of all nodes which are within the cluster radius. In the next step, the gray value matrices of each indicator are categorized with thresholds into three classes {1, 2, 3}. It would be also possible to keep the original values, but we chose categorization because the results are then easier and clearer to interpret for a human. In order to join the different layers and to achieve a final categorization of each individual grid cell, a function has to be defined enabling the combination of the different raster layers. This can be done with the weighted sum of all region indicators for each grid cell. In the example we combined the two layers with w1 = 0.5 and w2 = 0.5. The indicators of layer 1 and layer 2 have the values {1, 2, 3}. Both layers are combined with the weights 0.5. The possible values of the final layer are then {1, 1.5, 2, 2.5, 3}. Therefore the result is a final layer with five different classes. The values have no dimension and could be also normalized to the interval [0, 1]. In the following example we use vector data from the Geographic Data Files (GDF) in order to derive regions of different degrees of urbanity. GDF is an international standard for the modeling and exchange of road network data (ISO14825, 2004). GDF data are captured in a scale of approximately 1:25,000. The test area is shown in Fig. 9(a). Fig. 9(b) shows the grid cells superimposed on the vector data. Fig. 9(c) shows the result of calculating the rectangularity. Dark gray values stand for high rectangularity and bright gray values for low rectangularity. Fig. 9(d) shows the result of the calculation of the node density. Bright gray values indicate high node density whereas dark gray values indicate low node density. The result of this categorization is shown in Fig. 9(e) for the rectangularity and in Fig. 9(f) for the node density. Bright gray values stand for high urbanity and dark gray values for low urbanity. The evaluation of the point density leads to very good results whereas the result of the evaluation of the rectangularity is more diffuse. The reason for this is that the assumption, that in innercity areas more irregular, non-orthogonal roads can be found, is only partly true. In fact, in some inner-city areas we can find large areas with orthogonal roads and in some rural areas we can find large areas with non-rectangular roads. However, we use this layer for the further classification in order to show the combination of different indicators. These indicators must not come necessarily from the evaluation of grid cells. For example they can also come from a multispectral classification because the

526

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

a

boundaries. The main problem is to identify the fuzziness of the boundary which is different at different positions of the boundary. In future work we want to investigate this problem more in detail.

b

c

d

e

f

g

h

Fig. 9. Process chain: (a) input data (b) grid cells (c) rectangularity (d) node density (e) categorization of rectangularity (f) categorization of node density (g) combination of the measures (h) final result superimposed with input data.

result of multispectral classification is also a layer that contains grid cells with different classes that can be combined with other classes in the same way. Fig. 9(g) shows the combination of the two raster layers and Fig. 9(h) the final result superimposed with the vector input data. It can be seen that the detected urban areas are plausible. An evaluation of the result with ground truth data is difficult because of the fuzzy boundary of the regions. In Glemser and Fritsch (1998) it has been shown that objects with fuzzy boundaries are collected very different by different operators. Therefore we cannot compare directly fuzzy objects with objects with sharp

(a) Street data.

(b) Cluster size 50 m.

4.2.1. Influence of the parameters In this section it is shown in examples how the cluster size and the cluster radius influence the classification result. In the following examples we use only the node density indicator for the classification because it leads to clearer results as the combination of the node density and the rectangularity indicator as already discussed above. Fig. 10 shows the influence of the cluster size to the classification result. Fig. 10(a) shows the vector input data and Fig. 10(b)–(d) show the classification results with cluster size 50, 100 and 150 m. In all examples the cluster radius is 150 m and the thresholds for the categorization are threshold_low = 5 and threshold_high = 20. The cluster size has only an effect to the resolution of the result but not to the segmentation into different clusters. Therefore this is a very robust parameter that can be set to a wide range of values and does not require careful selection. Fig. 11 shows the influence of the cluster radius to the classification result. Fig. 11(b)–(d) show the classification results with cluster radius 150, 300 and 450 m. In all examples the cluster size is 100 m. The thresholds for the categorization have to be adapted to the different cluster radii because the number of topological nodes in a small area of influence is naturally smaller as in a large area of influence. Therefore we adapt the thresholds depending on the size of the area of influence and round them to the next integer value, as it is shown in Table 3. The adaptation of the thresholds has only to be done for indicators which are sensitive to the area of influence. For example the rectangularity indicator is always a value between 0 and 90 and therefore must not be adapted to different cluster radii. Increasing the cluster radius leads to a smoothing of the result. Therefore the cluster radius influences the scale of the result. The larger the cluster radius the smaller is the scale. Another effect is that if the cluster radius is much larger than the cluster size, the form of the clusters become more round, because the area of influence is calculated as a circle around the center of each grid cell. Again, this is robust parameter and does also not require careful selection. The influence of the thresholds to the classification result is shown in Fig. 12. The optimal thresholds were determined by a visual inspection of different test runs. Fig. 12(b) shows the classification result with optimal thresholds, Fig. 12(c) with higher thresholds and Fig. 12(d) with lower thresholds. A small increase or decrease of these thresholds has already significant effects on the result of the classification. Therefore the thresholds are the most sensitive parameters in the classification process and must

(c) Cluster size 100 m.

Fig. 10. Influence of the cluster size.

(d) Cluster size 200 m.

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

(a) Street data.

(b) Cluster radius 150 m.

(c) Cluster radius 300 m.

527

(d) Cluster size 450 m.

Fig. 11. Influence of the cluster radius. Table 3 Adapting of the thresholds for different cluster radii. Cluster radius

Area of influence = π ∗ cluster_radius2

Threshold low

Threshold high

150 300 = 2 ∗ 150 450 = 3 ∗ 150

70 685 282 743 ≈ 2 ∗ 2 ∗ 70 685 636 172 ≈ 3 ∗ 3 ∗ 70 685

5 20 = 2 ∗ 2 ∗ 5 45 = 3 ∗ 3 ∗ 5

20 80 = 2 ∗ 2 ∗ 20 180 = 3∗3∗20

(a) Street data.

(b) Low = 5, high = 20.

(c) Low = 10, high = 25.

(d) Low = 2, high = 15.

Fig. 12. Influence of the thresholds.

be determined carefully. Alternatively, these parameters can be determined semi-automatically with a supervised classification. In contrast the other two parameters node density and rectangularity are very robust. A detailed description of this approach can be found in Walter (2008). 5. Discussions The two examples show possible approaches for automatic map interpretation. They are straightforward and can be easily implemented. Nevertheless, a lot of research has still to be done. The most important task is to evaluate the quality of the results. Even if we could realize very high recognition rates, data which is interpreted wrongly or not certain has to be detected automatically. Measures have to be defined that describe the quality of the map interpretation. This enables the description of the quality of analyses that use the interpreted data as an input. Also it gives the possibility for a human, to control and revise the results of an automatic interpretation without looking at all the data. Furthermore methods are needed to describe the interpreted information with a formal language in order to enable a further automatic processing of this information. This can be done by using techniques from the Semantic Web. In the future we want to combine map interpretation techniques with the technologies from the Semantic Web. Furthermore we think that also the information about the quality of the interpretation has to be described with a formal language similar as the information itself. Therefore, the integration of quality measures is one of our major research tasks in the future.

6. Conclusions ‘‘Traditional spatial statistical and spatial analytical methods were developed in an era when data collection was expensive and computational power was weak’’ (Miller, 2007). This situation has changed. New sensor technologies, from high resolution multispectral satellites to terrestrial laser scanner, and new data collection methods — especially Web 2.0 mapping — provide spatial data in a fast way. Furthermore, computer power and storage capacity are mostly not a problem any more. And this development is still going on. Therefore, data collection has become cheap (or at least cheaper) and computational power is high and still increasing. In order to make full use of the potential of the huge amount of digital spatial data we need an automatic interpretation method to enrich the data with semantic information. The ongoing developments in computer technologies enable the development of new methods for spatial analyses which were not possible earlier because of their complexity. Map interpretation is a relatively new research area which is influenced by many different disciplines: Image Interpretation, Artificial Intelligence, Data Mining, Cognitive Science, Cartography, Psychology, Gestalt Theory, Pattern Recognition, Knowledge Discovery, Graph Theory and Clustering. In summary, map interpretation is something which can be described with ‘‘inverse cartographic engineering’’ because cartography is the practice of making maps in order to communicate spatial information effectively whereas map interpretation is the inverse task to translate and to model this information.

528

V. Walter, F. Luo / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 519–528

References Agarwal, A., Skupin, A. (Eds.), 2008. Self-Organizing Maps: Applications in Geographic Information Science. Wiley & Sons, West Sussex, England. Anders, K.H., 2003. A hierarchical graph-clustering approach to find groups of objects. ICA commission on map generalization. In: Technical Paper at the Fifth Workshop on Progress in Automated Map Generalization. IGN, Paris (Published on CD-ROM). http://www.geo.unizh.ch/ICA/docs/paris2003/papers03.html (accessed 11.01.11). Anders, K.H., Fritsch, D., 1996. Automatic interpretation of digital maps for data revision. International Archives of Photogrammetry and Remote Sensing 31 (Part B4), 90–94. Buccella, A., Cechich, A., Fillottrani, P., 2009. Ontology-driven geographic information integration: a survey of current approaches. Computers & Geosciences 35 (4), 710–723. Butenuth, M., Gösseln, G., Tiedge, M., Heipke, C., Lipeck, U., Sester, M., 2007. Integration of heterogenous geospatial data in a federated database. ISPRS Journal of Photogrammetry and Remote Sensing 62 (5), 328–346. Chen, H., Walter, V., 2009. Quality inspection and quality improvement of large spatial datasets. In: GSDI 11 World Conference: Spatial Data Infrastructure Convergence: Building SDI Bridges to Address Global Challenges (Published on CD-ROM). Dangermond, J., 2008. GIS and the GeoWeb. ArcNews. 30 (2) http://www.esri.com/news/arcnews/summer08articles/gis-andgeoweb.html (accessed 11.01.11). Feigenbaum, L., Herman, I., Hongsermeier, T., Neumann, E., Stephens, S., 2007. The Semantic Web in action. Scientific American 297 (6), 90–97. Filippovska, Y., Walter, V., Fritsch, D., 2008. Quality evaluation of generalization algorithms. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 37 (Part B2), 799–804. Fischer, F., 2008. Collaborative mapping—how wikinomics is manifest in the geoinformation economy. Geoinformatics 11 (2), 28–31. Frawley, W., Piatetsky-Shapiro, G., Matheus, C., 1991. Knowledge discovery in databases: an overview. In: Piatetsky-Shapiro, G., Frawley, W. (Eds.), Knowledge Discovery in Databases. AAAI, MIT Press, Menlo Park, CA, pp. 1–27. Glemser, M., Fritsch, D., 1998. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 32 (Part B4), 180–187. Google. 2008. Official Google blog: Google audio indexing now on Google labs: http://googleblog.blogspot.com/2008/09/google-audio-indexing-now-ongoogle.html (accessed 11.01.11). Graeff, B., Carosio, A., 2002. Automatic interpretation of raster-based topographic maps by means of queries. In: FIG XXII International Congress. Washington, DC (Published on CD-ROM). http://www.fig.net/pub/fig_2002/Ts310/TS3_10_graeff_carosio.pdf (accessed 11.01.11). Heinzle, F., Anders, K.H., Sester, M., 2007. Automatic detection of pattern in road networks—methods and evaluation. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 36 (Part 4/W45), (CD-ROM). http://tiny.cc/FEyb7 (accessed 11.01.11). Heinzle, F., Sester, M., 2004. Derivation of implicit information from spatial data sets with data mining. International Archives of Photogrammetry and Remote Sensing 35 (Part B4), 335–340. Heinzle, F., Sester, M., Anders, K.H., 2005. Graph-based approach for recognition of patterns and implicit information in road networks. In: Proceedings of 22nd International Cartographic Conference. La Coruña, Spain, 9–16 July (Published on CD-ROM). ISO14825, 2004. GDF—Geographic Data Files—Version 4. Berlin, Beuth Verlag. Jones, C., Purves, R., Clough, P., Joho, H., 2008. Modeling vague places with knowledge from the Web. International Journal of Geographical Information Science 22 (10), 1045–1065. Kohonen, T., 1982. Clustering, taxonomy, and topological maps of patterns. In: Proc. 6ICPR, Int. Conf. on Pattern Recognition. IEEE Computer Soc. Press, Washington, DC, pp. 114–128. Kuhn, W., 2002. Modeling the Semantics of Geographic Categories Through Conceptual Integration. In: Lecture Notes in Computer Science, vol. 2478. Springer, Berlin, Heidelberg, pp. 108–118.

Lüscher, P., Weiberl, R., Mackaness, A., 2008. Where is the terraced house? On the use of ontologies for recognition of urban concepts in cartographic databases. In: Ruas, A., Christopher, G. (Eds.), Headway in Spatial Data Handling. In: Lecture Notes in Geoinformation and Cartography, Springer, Berlin, Heidelberg, pp. 449–466. Mackaness, W., Edwards, G., 2002. The importance of modelling pattern and structure in automated map generalisation. In: Proceedings of Joint Workshop on Multi-Scale Representations of Spatial Data. Ottawa, Canada (Published on CD-ROM). http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.91. 1311&rep=rep1&type=pdf (accessed 11.01.11). Malerba, D., Esposito, F., Lanza, A., Lisi, F., Appice, A., 2003. Empowering a GIS with inductive learning capabilities: the case of INGENS. Computers, Environment and Urban Systems 27 (3), 265–281. Miller, H.J., 2007. Geographic data mining and knowledge discovery. In: Wilson, J.P., Fotheringham, A.S. (Eds.), Handbook of Geographic Information Science. WileyBlackwell, pp. 352–366. Miller, H.J., Han, J. (Eds.), 2009. Geographic Data Mining and Knowledge Discovery, second ed. Taylor and Francis. OpenStreetMap, 2010. OpenStreetMap: http://www.openstreetmap.org/ (accessed 11.01.11). Schleinkofer, M.F., 2007. Wissensbasierte Unterstützung zur Erstellung von Produktmodellen im Baubestand. Dissertation. Technische Universität München. Seo, S., O’Hara, C.G., 2009. Quality assessment of linear data. International Journal of Geographical Information Science 23 (12), 1503–1525. Sester, M., 2000. Knowledge acquisition for the automatic interpretation of spatial data. International Journal for Geographical Information Science 14 (1), 1–24. Sezgin, T.M., Davis, R., 2005. HMM-based efficient sketch recognition. In: Proceedings of the International Conference on Intelligent User Interfaces. IUI’05. ACM Press, pp. 281–283. Steinhauer, J.H., Wiese, T., Freksa, C., Barkowsky, T., 2001. Recognition of abstract regions in cartographic maps. In: Proceedings of the International Conference on Spatial Information Theory: Foundations of Geographic Information Science. pp. 306–321. Vailaya, A., Figueiredo, M., Jain, A., Zhang, H., 2001. Image classification for contentbased indexing. IEEE Transactions on Image Processing 10 (1), 117–129. Viglino, J.M., Pierrot-Deseilligny, M., 2003. A vector approach for automatic interpretation of the french cadatral map. In: Proceedings of the Seventh International Conference on Document Analysis and Recognition. ICDAR’03, pp. 304–309. Volz, S., Walter, V., 2006. Linking different geospatial databases by explicit relations. Geoinformation Science Journal 6 (1), 41–49. Walter, V., 2004. Object-wise classification of remote sensing data for change detection. In: Heipke, C., Pakzad, K., Willrich, F. (Eds.), Integration of Geodata and Imagery for Automated Refinement and Update of Spatial Databases. ISPRS Journal of Photogrammetry and Remote Sensing 58 (3–4), 225–238. Walter, V., 2008. Automatic interpretation of vector databases with a raster-based algorithm. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 37 (Part B2), 175–181. Walter, V., Fritsch, D., 1999. Matching spatial data sets: a statistical approach. International Journal of Geographical Information Science 13 (5), 445–473. Weindorf, M., 2002. Structure based interpretation of unstructured vector maps. In: Blostein, D., Kwon, Y. (Eds.), GREC 2002. In: LNCS, vol. 2390. Springer, Berlin, Heidelberg, pp. 190–199. Wiemann, S., Bernard, B., 2010. Conflation services within spatial data infrastructures. In: 13th AGILE International Conference on Geographic Information Science 2010. Portugal. http://agile2010.dsi.uminho.pt/pen/ShortPapers_ PDF%5C134_DOC.pdf (accessed 11.01.11). Wuersch, M., Egenhofer, M.J., 2008. Perceptual sketch interpretation. In: Ruas, A., Gold, C. (Eds.), The 13th International Symposium on Spatial Data Handling. SDH 2008. Springer, Montpellier, France, pp. 19–38. Zhang, Q., 2004. Modeling structure and patterns in road network generalisation. In: Workshop on Generalisation and Multiple Representation. Leicester, UK. http://ica.ign.fr/Leicester/paper/Zhang-v1-ICAWorkshop.pdf (accessed 11.01.11). Zhang, M., Meng, L., 2006. Implementation of a generic road-matching approach for the integration of postal data. In: Proceedings of 1st ICA Workshop on Geospatial Analysis and Modeling. Vienna, Austria, pp. 141–154.